Fuzzy propositions weighted by veracities or how to relate fuzzy logic and probability theory for segmentation of ultrasound images

In medical imaging, and more generally in medical information, researches go towards fusion systems. Nowadays, the steps of information source definition, the pertinent data extraction and the fusion need to be conducted as a whole. In this work, our interest is related to the esophagus wall segmentation from ultrasound images sequences. We aim to elaborate a general methodology of data mining that coherently links works on data selection and fusion architectures, in order to extract useful information from raw data and to integrate efficiently the physician a prior. In the presented method, based on fuzzy logic, some fuzzy propositions are defined using physicians a prior knowledge. The use of probabilistic distributions, estimated thanks to a learning base of pathologic and non-pathologic cases, enables the veracity of these propositions to be qualified. This promising idea enables information to be managed through the consideration of both information imprecision and uncertainty. In the same time, the obtained benefit, when a prior knowledge source is injected in a fusion based decision system, can be quantified. By considering that, the fuzzyfication stage is optimized relatively to a given criteria using a genetic algorithm. By this manner, fuzzy sets corresponding to the physicians ambiguous a prior are defined objectively. At this level, we successively compare performances obtained when fuzzy functions are defined empirically and when they are optimized. We conclude this paper with the first results on esophagus wall segmentation and outline some further works.